On Uncertainty Calibration and Selective Generation in Probabilistic Neural Summarization: A Benchmark Study